When Replanning Becomes the Bottleneck: Budgeted Replanning for Embodied Agents

Feiyang You; Ningxin Su; Shuaijun Liu; Xingwei Chen

REVIEW 3 major objections 5 minor 15 references

When Replanning Becomes the Bottleneck: Budgeted Replanning for Embodied Agents

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Long-horizon embodied agents fail when replanning-call latency grows, not when plan quality drops: a budgeted controller plus context pruning restores schedulability while preserving success.

desk verdict A genuinely useful systems contribution to replanning accounting and budgeting, but the headline 80% success claim rests on an unspecified harder setting and undisclosed controller thresholds; it deserves a serious referee but needs major revision. read the letter →

arxiv 2608.01428 v1 pith:KTMDEWAG submitted 2026-08-02 cs.RO cs.AIcs.LG

Shuaijun Liu , Feiyang You , Xingwei Chen , Ningxin Su This is my paper

classification cs.ROcs.AIcs.LG

keywords budgetedreplanningembodiedagentstokenpruningtaillatencySLOviolationratecontextgrowthmulti-agentcoordinationLLM/VLAplanning

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper's central claim is that in long-horizon embodied tasks the bottleneck is the replanning call itself: as prompts accumulate observations, failure traces, and multi-agent messages, transformer latency grows heavy-tailed and misses real-time deadlines even while task success stays at 100%. To address this, the paper proposes BRACE, a controller that treats every replanning trigger as a budget decision—whether to replan, in which mode, and with which token and latency budgets—and E-RECAP, a trained progressive token pruner that compresses the accumulated context while preserving head and tail tokens. Across Meta Habitat, RoboFactory, and AirSim, the combination cuts replanning-call tokens by 62–92% and reduces violations of the per-call latency service-level objective (SLO) from 85.5–100% to 4.7–50% in regimes where success was already saturated. In a harder RoboFactory setting where open-loop, frozen-plan, and No BRACE all score 0%, BRACE + E-RECAP reaches 80% success with 4.6% SLO violations, the paper's key evidence that tail-aware per-call budgeting can preserve recovery while restoring schedulability.

What carries the argument

Two objects carry the argument. The first is BRACE, a closed-loop controller that converts each trigger signal $\tau_t$ into a triple decision: whether to invoke the planner, which replanning mode to use, and a token budget $B_t$ with a latency SLO. Its stabilizing core is the cooldown/commit gate of Eq. (1), which forces a deterministic minimum spacing between replanning calls and between plan replacements; failure-aware overrides can relax the gate after repeated failures. The second is E-RECAP, a progressive token-pruning module trained to predict token utility from hidden states: at pruning layer $l$ it sets $N_{l+1} = \lfloor r_l N_l \rfloor$, always keeps a fixed head set and a tail wi

What would settle it

Run the RoboFactory harder setting with the same budget and pruning but with the controller's cooldown/commit/deadlock values set to configurations that the paper's own proxy sweep shows failing (e.g., $\delta=3$, $\omega=2$, $w=5$ gives 3.3% success). If success drops toward that range, the headline is carried by the knob settings rather than by budgeting-plus-pruning. Conversely, if removing E-RECAP while keeping the BRACE gate leaves P95 and the SLO violation rate nearly unchanged, then pruning, not budgeting, is doing the schedulability work.

Watch

Extended reading notes

Core claim

The central claim is that replanning-call latency, not plan correctness, is the bottleneck under context growth. As accumulated prompts grow, transformer attention cost rises superlinearly and call latency becomes heavy-tailed while task success stays flat. BRACE converts each trigger into a budgeted decision via the gate $u_t = \mathbb{I}[\tau_t \wedge (\Delta_t \ge \delta) \wedge (\kappa_t \ge \omega)]$ and per-call token/latency budgets; E-RECAP prunes progressively, keeping head/tail tokens and filling the rest by predicted utility ($N_{l+1} = \lfloor r_l N_l \rfloor$). Across Habitat, RoboFactory, and AirSim $K{=}8$, No BRACE retains 100% success while violating the SLO on 85.5–100% of

Load-bearing premise

The load-bearing premise is that the 'harder' RoboFactory setting differs from the saturated one only by task difficulty, and that the controller thresholds (cooldown, commit, failure-aware overrides) were chosen by a fixed rule rather than tuned to favor BRACE + E-RECAP; the paper reports neither the harder-setting recipe nor the deployed threshold values, and its own proxy sweep shows success ranging from 98.3% to 3.3% as nearby settings change.

Editorial extensions

If this is right

Task success alone can be a misleading closed-loop metric: a planner can score 100% success while missing its latency deadline on most replanning calls, so tail percentiles and SLO violation rates should be reported at call granularity.
Reducing replanning tokens is necessary but not sufficient: at the same token budget, recency truncation, random truncation, and structured summaries behave differently, so context-reduction quality must be compared under matched budgets.
Controller-side stabilization (cooldown, commit windows, failure-aware overrides) reduces replanning churn and coordination wait; pruning alone does not remove the feedback loops caused by repeated failures or coordination hazards.
Budgeted replanning composes with efficiency modules such as pruning and retrieval, and the phase accounting makes the overhead of each module auditable instead of attributing all latency to the planner.
In the harder RoboFactory regime, the same interface preserves recovery behavior where open-loop, frozen-plan, and No BRACE all fail, showing that budgeting can matter for task quality, not only for schedulability.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the bottleneck diagnosis generalizes beyond the three platforms, the same controller/pruner interface could apply to other long-context agent loops—code agents, web agents, or human-in-the-loop systems—where latency SLOs and context growth co-occur; this is an extension the paper does not test.
The strong dependence of success on cooldown/commit/deadlock settings in the paper's own proxy sweep suggests an automatic tuning rule: given a latency budget and an SLO violation allowance, choose the smallest cooldown/commit windows that stabilize churn, turning the controller from a fixed knob into a self-calibrating policy.
E-RECAP's head/tail preservation rule implies that prompt-structure-aware pruning, not just token-saliency pruning, is the operative mechanism; a testable corollary is that pruning schemes that protect task specification and recent observations should transfer across planner backbones better than pure saliency pruning.
Using SLO violation rate as a first-class metric would change how planners and VLA models are compared: two models with identical success can be separated by their call-latency distributions. That reframing is implicit in the paper's accounting but not stated as a selection criterion.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The genuinely new thing here is the call-level budgeting and accounting formalism: per-replanning-call token budgets and latency SLOs, phase-level attribution, and budget-matched baselines. That is a real contribution that systems-minded robotics and LLM researchers will find useful, and it is presented cleanly. The paper also earns credit for being unusually transparent in spots: it shows recency truncation beating E-RECAP on SLO violations in Habitat, concedes that the anti-churn property is a deterministic consequence of the gate, and publishes a controller sweep that exposes how sensitive the results are to the knobs.

The soft spot is the central success claim. The "harder" RoboFactory setting that produces the 80% headline number is never specified relative to the saturated Pass-Shoe condition; the controller thresholds used in the main experiments are not reported; and the one disclosed sweep (Appendix Table 30) shows success swinging from 98.3% to 3.3% across nearby settings. Combined with 10-episode cells and no error bars, the headline figure is not reproducible or distinguishable from a favorable knob choice. There is also an internal tension: the harder setting shows only 4.6% SLO violations with P95=247ms, while the saturated Pass-Shoe setting in Table 2 shows 50% violations with P95=1213ms. Since the harder setting is defined nowhere, that combination cannot be interpreted. A smaller issue: the AirSim near-miss rate worsens from 24.7 to 29.2 per episode under BRACE+E-RECAP, and that goes without discussion. No code, data, or commit hash is released.

What holds up is the saturated-setting claim: the paper does demonstrate that success can sit at 100% while SLO violations persist at 85.5–100%, and that budgeted replanning with a context pruner can cut those violations substantially. The Habitat matched-budget results also show that a trivial recency truncation achieves 0.0% violations, which actually strengthens the point that the bottleneck is schedulability and budgeting, not the specific learned pruner.

The paper is for researchers building LLM/VLA-driven embodied systems with real-time constraints, and for benchmark designers who want to move beyond success-only reporting. The accounting and control-loop contribution holds up on its own, independent of the E-RECAP module results.

Recommendation: send this to reviewers. The problem is real and the formalism is a useful primitive. But the authors should be required to specify the harder setting, report the deployed controller thresholds, add variance estimates or more episodes, and release code and configuration before the headline claim can be accepted.

Referee Report

3 major / 5 minor

Summary. Embodied agents that replan with LLMs accumulate long contexts, so each replanning call develops heavy-tail latency and can miss real-time SLOs even when task success is high. The paper proposes BRACE, a controller that decides whether to replan, selects a replanning mode, and allocates token and latency budgets, with cooldown/commit windows and failure overrides. It also introduces E-RECAP, a progressive token-pruning module that predicts token utility and preserves head/tail tokens. Experiments on Habitat, RoboFactory, and AirSim report large reductions in replanning tokens (62–92%) and SLO violations (85.5–100% to 4.7–50%) in success-saturated regimes, plus a 'harder' RoboFactory setting where BRACE+E-RECAP reaches 80% success while open-loop, frozen-plan, and No BRACE fail. A focused real-robot study on PICKFRUIT and PUSHT is also included.

Significance. If substantiated, the paper identifies a real and under-reported failure mode: success can be saturated while the replanning call path repeatedly misses deadlines, and this can be addressed by a budget controller plus context pruning. The contribution is strengthened by call-level phase accounting, budget-matched baselines, and the inclusion of no-replanning and frozen-plan conditions. The paper is also transparent about some limitations: E-RECAP's benefit over recency truncation in Habitat is small, and the anti-churn property is explicitly a deterministic consequence of the gate. However, the headline 'harder setting' is not specified, and the controller thresholds used in all main experiments are not reported; given the paper's own sensitivity sweep, the evidence is currently insufficient to support the general claim.

major comments (3)

[Section 4.4, Table 6; Appendix Table 9] The RoboFactory 'harder setting' that produces the 80% success headline is never defined. Appendix Table 9 only labels the rows 'harder Pass-Shoe' and says the configuration 'exposes the necessity of replanning'; it does not state what changes relative to the saturated Pass-Shoe of Table 2 (task variant, noise, context-growth rate, token budget, SLO, or episode count). The internal inconsistency is severe: Table 2 reports P95=1213 ms and 50% SLO violations for the saturated Pass-Shoe, while Table 6/9 reports P95=247 ms and 4.6% violations for the 'harder' setting, implying a smaller context or a different SLO regime. Without a definition of this setting, the 80% success claim cannot be reproduced, compared to baselines, or interpreted as evidence for the method rather than for a favorable configuration.
[Section 3.2, Eq. (1); Appendix Table 30] The values of the cooldown threshold δ, the commit threshold ω, and the failure-aware override rules are not reported for any of the main experiments (Tables 2, 6, 7). The only disclosed sweep of these knobs, Appendix Table 30, shows success ranging from 98.3% to 3.3% when δ, ω, and the deadlock window w move through nearby settings. Because the deployed settings are unspecified, the headline results are not reproducible and may reflect a favorable parameter choice. This is load-bearing: the paper's main claim is that BRACE's controller stabilizes replanning; without the controller configuration, that claim is unverifiable.
[Tables 2 and 6; Section 4.3] The headline tables report point estimates with no confidence intervals, standard errors, or statistical tests, and the episode counts are small (10–30 per condition). Given that the paper's contribution is quantitative (SLO violation rates dropping from 85.5–100% to 4.7–50%), the absence of any uncertainty quantification makes it impossible to assess whether differences like 50% vs 4.6% or 0% vs 80% are stable. A per-seed breakdown or at least a bootstrap interval for the key cells should be provided.

minor comments (5)

[Section 4.4] The section is titled 'Ablations' but Table 6 contains open-loop and harder-setting comparisons, which are additional baselines rather than ablations; consider renaming the section or separating the two.
[Appendix A.1] Appendix Table 9 appears in a subsection titled 'Pruning configuration' but actually reports the harder-setting results; retitle or move the table to the experimental section where it is used.
[Eq. (6)] The symbol T is used for the number of replanning calls in Eq. (6), but T (or t) is also used for the controller step throughout Section 3; use a different symbol (e.g., N_calls) to avoid ambiguity.
[Table 2] The column header 'Ep' is not defined; specify that it denotes the number of episodes per condition.
[Eq. (3)] In Eq. (3), n_{i,a} is not defined in the text; clarify that it denotes the token contribution of agent a between controller steps i−1 and i.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: BRACE/E-RECAP's equations are definitions and the empirical claims are measured against external benchmarks; the main concerns are reproducibility, not derivation-by-construction.

full rationale

The paper's derivation chain is not circular. Equations (1)-(6) define the stability gate, budget selection, context-growth decomposition, pruning schedule, kept-index set, and SLO violation rate; none of these definitions presupposes the headline reductions. The anti-churn property is explicitly conceded to be a deterministic consequence of Eq. (1), not a theorem-level result. E-RECAP's token-utility predictor is trained on instruction-following and embodied auxiliary data (Dolly, Alpaca, Self-Instruct, ALFRED, TEACh, BabyAI) and evaluated on MP3D PointNav/ObjectNav plus RoboFactory/AirSim, so the reported success/SLO outcomes are not fitted inputs renamed as predictions. The token-reduction percentages are set by the keep-ratio schedule (Eq. 4), but the paper presents them as implementation parameters, not as discovered predictions. There are no self-citations used as load-bearing evidence and no imported uniqueness theorem. The genuine weaknesses--the unspecified 'harder Pass-Shoe' setting in Table 6 / Appendix Table 9 and the unreported controller thresholds (delta, omega, w) that Appendix Table 30 shows to be highly sensitive--are reproducibility/configuration-reporting concerns, not circularity.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The headline numbers are anchored by author-chosen design points (keep ratio, per-domain budgets and SLOs, undisclosed controller thresholds, unspecified harder-setting recipe) rather than derived from task structure; the paper's closest-to-circular element is that the reported SLO gains are defined against thresholds the authors select. No physical or formal entities are invented; BRACE and E-RECAP are engineering modules, so the graviton problem does not apply.

free parameters (6)

E-RECAP keep ratio r = 0.7
Headline token-reduction and speedup results use r=0.7 (Table 5) after sweeping r in {0.9,0.8,0.7,0.6,0.5} (Appendix Table 28, Figure 18); quality retention is fitted to this choice.
Per-domain token budgets Bt = 20 (Habitat), 128 (RoboFactory), N/A (AirSim)
Budget-matched baselines are forced to these hand-set budgets (Table 11); the token-count and SLO gains are measured against this author-chosen constraint.
Per-domain latency SLOs = 2500 ms (Habitat), 250 ms (RoboFactory), 2500 ms (AirSim)
The headline SLO violation rates are defined relative to these chosen deadlines; RoboFactory's 250 ms SLO keeps violation at 50% even for the best method, and no sensitivity sweep over SLO values is given.
Controller thresholds delta, omega on main platforms = undisclosed in main text
The stability gate (Eq. 1) depends on cooldown/commit thresholds that are unspecified for the main experiments; the proxy sweep (Table 30) shows success varying between 98.3% and 3.3% across nearby settings.
Head/tail windows and pruning schedule = head=4, tail=max(16, 0.1*N), layers {4,7,10,...}
Hand-set in Appendix Table 8; they determine how many tokens survive independent of the utility predictor, affecting the measured token budget and quality.
Training mixture weights and loss coefficients = 90/10 and 85/5/5 mixtures; lambda_1, lambda_2
Chosen via sweeps in Appendix Table 28; effect on quality is small (0-2.5% SPL), so this is a minor fitted knob.

assumptions (4)

domain assumption Transformer attention compute grows superlinearly in context length, approximated as O(N^2) (Section 3.3)
Motivates the entire pruning approach, but measured replanning speedups (2.09-2.64x for 71-76% token removal) are far below O(N^2) scaling, so the assumed bottleneck structure is only loosely matched.
domain assumption Replanning context is redundant enough that pruning to 20-128 tokens preserves the information needed for successful recovery (Section 3.3)
Tested empirically on the chosen tasks, and true only where success stays saturated; in the only non-saturated setting the pruned system also changes the controller, so the pruner's contribution is not isolated.
domain assumption Call-granularity SLO violation and P95 latency are the right measures of closed-loop stability (Eq. 6, Table 1)
The paper's novel metric framing presumes that deadline misses, not plan quality, dominate closed-loop behavior; success could mask violations only if the planner is otherwise adequate.
ad hoc to paper The unspecified RoboFactory 'harder setting' is a representative hard regime, not a configuration engineered to favor the method (Section 4.4, Table 6)
No definition of the harder change is given; this is the load-bearing regime for the 80% success claim.

how reviews work

0 comments

Cite this review

Pith. "Pith review of When Replanning Becomes the Bottleneck: Budgeted Replanning for Embodied Agents." pith.science (2026). https://pith.science/paper/KTMDEWAG

@misc{pith2026260801428,
  author       = {Pith},
  title        = {Pith review of: When Replanning Becomes the Bottleneck: Budgeted Replanning for Embodied Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KTMDEWAG}},
  note         = {Machine review of arXiv:2608.01428}
}

read the original abstract

Embodied agents replan frequently to recover from execution drift, partial observability, and coordination hazards, but each LLM-based replanning call can consume an accumulated textual context that grows over time and across agents. Once this context becomes large, replanning latency develops heavy tails and can miss real-time deadlines even when task success remains high, a failure mode that is hard to detect from average latency or success alone. We present BRACE, a controller that formulates replanning as a budgeted control loop by deciding whether to replan, selecting a replanning mode, and allocating an explicit token budget and latency service-level objective (SLO) while accounting for optional efficiency modules. As a reusable component, we introduce E-RECAP, a cost-aware progressive token pruning method that predicts token utility and prunes replanning contexts across transformer layers while preserving critical head and tail tokens. Across Meta Habitat, RoboFactory, and AirSim, BRACE with E-RECAP reduces replanning-call token counts by 62-92% and SLO violation rates from 85.5-100.0% to 4.7-50.0% in settings where task success is already saturated. In a harder RoboFactory setting where open-loop, frozen-plan, and No BRACE all fail, BRACE + E-RECAP reaches 80.0% success with 4.6% SLO violations, demonstrating that tail-aware per-call budgeting is effective across embodied platforms.

Figures

Figures reproduced from arXiv: 2608.01428 by the authors.

**Figure 1.** Qualitative motivating example on Meta Habitat. As replanning history accumulates, the context grows and produces tail-latency spikes, motivating budgeted context compression. retrieval, plan caching, and communication compression). To make composition comparable, BRACE instruments the replanning call path with phase accounting and audit logs. For each replanning call, we record tokens and latency for context compre… view at source ↗

**Figure 2.** Overview of BRACE: a closed-loop controller that determines whether to invoke replanning and selects the per-call token budget and latency SLO, with E-RECAP as a composable token-pruning module on the replanning call path. 3.2. BRACE Controller: When to Replan and How Much to Spend BRACE maps trigger signals and a small controller state to replanning decisions and per-call budgets. It has two responsibilities: (1) d… view at source ↗

**Figure 3.** E-RECAP module. A lightweight predictor scores context tokens using intermediate hidden states, and pruning is applied progressively at selected layers. The kept set includes fixed head and tail tokens plus top-scoring tokens, producing a shorter context that satisfies the planner-input budget. 3.4. Phase Accounting and Audit Logging BRACE instruments the replanning call path with phaselevel accounting. For each co… view at source ↗

Figures from the paper (15 more)

**Figure 4.** Figure 4: summarizes distribution-level tail behavior: (a) a CDF with the SLO threshold (Meta Habitat) and (b) a cross-platform SLO violation snapshot. Why it matters: tail percentiles and violation rates directly correspond to 2000 2200 2400 2600 2800 Replanning latency (ms) 0.…

**Figure 5.** Figure 5: Qualitative example on RoboFactory TakePhoto (multiagent), showing representative frames from the baseline and BRACE rollouts. Quantitative comparisons are reported in Table 2, and the corresponding ablations in [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

**Figure 6.** Figure 6: Qualitative comparison on AirSimNH. The baseline accumulates delayed replanning decisions within the same trigger window, whereas BRACE shortens the effective replanning path and completes the interaction with fewer deadline misses. The corresponding quantitative resul…

**Figure 7.** Figure 7: Real-robot PICKFRUIT rollout pair on a banana instance. The upper rollout uses BRACE and triggers replanning twice after an intermediate execution failure, subsequently recovering and completing the task; the lower rollout relies solely on the underlying LLM+VLA stack …

**Figure 8.** Figure 8: Diagnostic views of coordination wait, SLO violations, and token–latency tradeoffs across the RoboFactory and Habitat domains. wait time and SLO violations as additional stability signals. Appendix [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]

**Figure 9.** Figure 9: E-RECAP experimental setup overview. The diagram summarizes the replanning data flow, pruning module, and planner interface used for the Habitat acceleration experiments [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]

**Figure 10.** Figure 10: Qualitative extension on RoboSuite Two-Arm Handover. The contact sheet shows a successful coordinated manipulation rollout beyond the RoboFactory benchmark [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]

**Figure 11.** Figure 11: Qualitative comparison on RoboFactory Pass-Shoe between the baseline and BRACE+E-RECAP. The paired frames illustrate how BRACE stabilizes replanning and reduces coordination delay during the manipulation handoff [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]

**Figure 12.** Figure 12: Additional qualitative example on RoboFactory TakePhoto. The frame sequence depicts a successful multi-agent manipulation rollout that complements the Pass-Shoe hardersetting results. Main-paper [PITH_FULL_IMAGE:figures/full_fig_p017_12.png]

**Figure 13.** Figure 13: Full-rollout comparison on RoboSuite TWO-ARM PEG-IN-HOLE. The combined contact sheet contrasts a BRACE rollout that stabilizes replanning and successfully completes the task against the corresponding No-BRACE baseline, which repeatedly triggers replanning without con…

**Figure 14.** Figure 14: Training-data configuration overview for E-RECAP. The figure summarizes the data mixture used by the token-utility predictor and corresponds to the quantitative comparison in Table 28 [PITH_FULL_IMAGE:figures/full_fig_p018_14.png]

**Figure 15.** Figure 15: Visualization of the proxy controller sweep. Success and Deadlocks/ep are plotted as a function of the deadlock window w, with separate lines for different cooldown δ and commit window ω settings ( [PITH_FULL_IMAGE:figures/full_fig_p019_15.png]

**Figure 16.** Figure 16: Successful BRACE rollout on the LIBERO Kitchen Scene 3 Moka Pot task. The replanning interface generalizes to a kitchen manipulation scenario with a different visual and action distribution. Front Grasp Camera Overhead Camera “Take the banana on the left into the gree…

**Figure 17.** Figure 17: Qualitative example on the Microsoft AirSim LandscapeMountains scene (K=8 agents). Left: a baseline that continues to trigger replanning after collision- or disturbance-induced drift but does not apply BRACE’s controller-level budgeting and stabilization. Right: BRACE…

**Figure 18.** Figure 18: E-RECAP tradeoff between compression and performance under stronger pruning. The plot shows how latency and token savings change as the keep ratio decreases, alongside the resulting task-quality trend. median goal error decreases by 22.1% within 13.2s after the trigge…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

15 extracted references · 7 canonical work pages

[8]

Puig, X., Undersander, E., Szot, A., Cote, M

doi: 10.1609/aaai.v36i2.20097. Puig, X., Undersander, E., Szot, A., Cote, M. D., Yang, T.-Y ., Partsey, R., Desai, R., Clegg, A. W., Hlavac, M., Min, S. Y ., V ondruˇs, V ., Gervet, T., Berges, V .-P., Turner, J. M., Maksymets, O., Kira, Z., Kalakrishnan, M., Malik, J., Chaplot, D. S., Jain, U., Batra, D., Rai, A., and Mot- taghi, R. Habitat 3.0: a co-hab...

doi:10.1609/aaai.v36i2.20097
[9]

RoboFactory: exploring embodied agent collaboration with compositional constraints.arXiv preprint arXiv:2503.16408,

Qin, Y ., Kang, L., Song, X., Yin, Z., Liu, X., Liu, X., Zhang, R., and Bai, L. RoboFactory: exploring embodied agent collaboration with compositional constraints.arXiv preprint arXiv:2503.16408,

arXiv
[10]

Wang, M., Chu, J., Xie, S., Zang, X., Zhao, Y ., and Zhong, W

See also the CRFM release post at https://crfm.stanford.edu/2023/03/13/alpaca.html. Wang, M., Chu, J., Xie, S., Zang, X., Zhao, Y ., and Zhong, W. Put teacher in student’s shoes: cross-distillation for ultra-compact model compression framework. InACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD), 2025a. Wang, Y ., Kordi, Y ., Mishra, S., ...

work page 2023
[11]

doi: 10.18653/v1/2023.acl-long

doi:10.18653/v1/2023.acl-long 2023
[12]

Xu, S., Wang, Y ., Xia, C., Zhu, D., Huang, T., and Xu, C

doi: 10.1109/ICRA55743.2025.11128420. Xu, S., Wang, Y ., Xia, C., Zhu, D., Huang, T., and Xu, C. VLA-Cache: efficient vision-language-action manipula- tion via adaptive token caching. InAdvances in Neural Information Processing Systems,

arXiv 2025
[13]

Zandieh, A., Daliri, M., Hadian, M., and Mirrokni, V

doi: 10.24963/IJCAI.2025/766. Zandieh, A., Daliri, M., Hadian, M., and Mirrokni, V . Tur- boQuant: online vector quantization with near-optimal distortion rate. InInternational Conference on Learning Representations (ICLR),

work page doi:10.24963/ijcai.2025/766 2025
[14]

Table 13 provides the corresponding ObjectNav extension and shows that the same pruning regime remains effective beyond a single navigation task family. A.2. Additional Domain Tables The following tables provide additional per-domain results that complement the main-paper evidence snapshot by ex- posing stability and safety proxy metrics that are easy to ...

work page 2025
[16]

Training-data sources referenced in Table 28 include Dolly, Alpaca, and Self-Instruct (Conover et al., 2023; Taori et al., 2023; Wang et al., 2023)

that support the standalone pruning module evidence used throughout the BRACE paper. Training-data sources referenced in Table 28 include Dolly, Alpaca, and Self-Instruct (Conover et al., 2023; Taori et al., 2023; Wang et al., 2023). Figure 18 complements the tables by summarizing the compression– quality/efficiency tradeoff across pruning strengths. In T...

work page 2023

Show all 15 references

[20]

The key point is that pruning overhead remains small relative to planner time, so the main system gain comes from shorten- ing the planner input under the same call-path accounting

Habitat phase overhead.Table 16 separates E-RECAP- side overhead from planner latency on the Habitat slice. The key point is that pruning overhead remains small relative to planner time, so the main system gain comes from shorten- ing the planner input under the same call-path...

2026
[2017]

Chang, M., Chhablani, G., Clegg, A., Cote, M

doi: 10.1109/3DV .2017.00081. Chang, M., Chhablani, G., Clegg, A., Cote, M. D., Desai, R., Hlavac, M., Karashchuk, V ., Krantz, J., Mottaghi, R., Parashar, P., Patki, S., Prasad, I., Puig, X., Rai, A., Ramrakhya, R., Tran, D., Truong, J., Turner, J. M., Un- dersander, E., and ...

2017
[2022]

Reflective planning: vision-language models for multi- stage long-horizon robotic manipulation

Feng, Y ., Han, J., Yang, Z., Yue, X., Levine, S., and Luo, J. Reflective planning: vision-language models for multi- stage long-horizon robotic manipulation. In Lim, J., Song, S., and Park, H.-W. (eds.),Proceedings of The 9th Con- ference on Robot Learning, volume 305 ofProce...
[2023]

ToDRE: effective visual token pruning via token diversity and task relevance, 2025a

Li, D., Yang, Z., Zhang, X., Shao, L., and Lu, S. ToDRE: effective visual token pruning via token diversity and task relevance, 2025a. Li, Y ., Meng, Y ., Sun, Z., Ji, K., Tang, C., Fan, J., Ma, X., Xia, S., Wang, Z., and Zhu, W. SP-VLA: a joint model scheduling and token prun...
[2024]

TEACh: task-driven embodied agents that chat

Padmakumar, A., Thomason, J., Shrivastava, A., Lange, P., Narayan-Chen, A., Gella, S., Piramuthu, R., Tur, G., and Hakkani-Tur, D. TEACh: task-driven embodied agents that chat. InProceedings of the AAAI Conference on Artificial Intelligence (AAAI), volume 36, pp. 2017–2025,

2017
[2025]

J., Finn, C., and Liang, P

Kim, M. J., Finn, C., and Liang, P. Fine-tuning vision- language-action models: optimizing speed and success. arXiv preprint arXiv:2502.19645, 2025a. Kim, M. J., Pertsch, K., Karamcheti, S., Xiao, T., Balakr- ishna, A., Nair, S., Rafailov, R., Foster, E. P., Sanketi, P. R., Vu...

arXiv
[2026]

Goyal, A., Hadfield, H., Yang, X., Blukis, V ., and Ramos, F

Poster. Goyal, A., Hadfield, H., Yang, X., Blukis, V ., and Ramos, F. VLA-0: building state-of-the-art VLAs with zero modification.arXiv preprint arXiv:2510.13054,

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [8]

Puig, X., Undersander, E., Szot, A., Cote, M

doi: 10.1609/aaai.v36i2.20097. Puig, X., Undersander, E., Szot, A., Cote, M. D., Yang, T.-Y ., Partsey, R., Desai, R., Clegg, A. W., Hlavac, M., Min, S. Y ., V ondruˇs, V ., Gervet, T., Berges, V .-P., Turner, J. M., Maksymets, O., Kira, Z., Kalakrishnan, M., Malik, J., Chaplot, D. S., Jain, U., Batra, D., Rai, A., and Mot- taghi, R. Habitat 3.0: a co-hab...

doi:10.1609/aaai.v36i2.20097

[2] [9]

RoboFactory: exploring embodied agent collaboration with compositional constraints.arXiv preprint arXiv:2503.16408,

Qin, Y ., Kang, L., Song, X., Yin, Z., Liu, X., Liu, X., Zhang, R., and Bai, L. RoboFactory: exploring embodied agent collaboration with compositional constraints.arXiv preprint arXiv:2503.16408,

arXiv

[3] [10]

Wang, M., Chu, J., Xie, S., Zang, X., Zhao, Y ., and Zhong, W

See also the CRFM release post at https://crfm.stanford.edu/2023/03/13/alpaca.html. Wang, M., Chu, J., Xie, S., Zang, X., Zhao, Y ., and Zhong, W. Put teacher in student’s shoes: cross-distillation for ultra-compact model compression framework. InACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD), 2025a. Wang, Y ., Kordi, Y ., Mishra, S., ...

work page 2023

[4] [11]

doi: 10.18653/v1/2023.acl-long

doi:10.18653/v1/2023.acl-long 2023

[5] [12]

Xu, S., Wang, Y ., Xia, C., Zhu, D., Huang, T., and Xu, C

doi: 10.1109/ICRA55743.2025.11128420. Xu, S., Wang, Y ., Xia, C., Zhu, D., Huang, T., and Xu, C. VLA-Cache: efficient vision-language-action manipula- tion via adaptive token caching. InAdvances in Neural Information Processing Systems,

arXiv 2025

[6] [13]

Zandieh, A., Daliri, M., Hadian, M., and Mirrokni, V

doi: 10.24963/IJCAI.2025/766. Zandieh, A., Daliri, M., Hadian, M., and Mirrokni, V . Tur- boQuant: online vector quantization with near-optimal distortion rate. InInternational Conference on Learning Representations (ICLR),

work page doi:10.24963/ijcai.2025/766 2025

[7] [14]

Table 13 provides the corresponding ObjectNav extension and shows that the same pruning regime remains effective beyond a single navigation task family. A.2. Additional Domain Tables The following tables provide additional per-domain results that complement the main-paper evidence snapshot by ex- posing stability and safety proxy metrics that are easy to ...

work page 2025

[8] [16]

Training-data sources referenced in Table 28 include Dolly, Alpaca, and Self-Instruct (Conover et al., 2023; Taori et al., 2023; Wang et al., 2023)

that support the standalone pruning module evidence used throughout the BRACE paper. Training-data sources referenced in Table 28 include Dolly, Alpaca, and Self-Instruct (Conover et al., 2023; Taori et al., 2023; Wang et al., 2023). Figure 18 complements the tables by summarizing the compression– quality/efficiency tradeoff across pruning strengths. In T...

work page 2023